OpenAI announced the release of its next-generation artificial intelligence model line, GPT-5.6, arriving just two months after the debut of its predecessor, GPT-5.5.

The rollout departs from the tech company’s previous direct-to-consumer deployment strategies seen with ChatGPT and Codex, entering instead into a restricted phase coordinated with United States officials.

According to company statements, the GPT-5.6 generation introduces a celestial-themed naming convention designed to separate a model’s foundational generation from its specific performance tier.

The new framework establishes three distinct versions aimed at balancing capability, speed, and operational cost for developers and the broader public.

Sol serves as the flagship model of the new lineup, designed to handle the most intensive computation and reasoning tasks.

Terra represents the intermediate tier, positioned as a balanced option for standard daily workflows, while Luna fills the final slot as a faster, more affordable option.

Company documents indicate that while the 5.6 designation tracks the overall generational leap, these individual tier names will persist as independent, durable capability brackets that can receive updates on separate timelines.

The immediate availability of the models remains heavily restricted due to an ongoing arrangement between the artificial intelligence firm and federal regulators.

OpenAI confirmed it previewed the technical capabilities and deployment plans for the GPT-5.6 series with the U.S. government prior to making any public announcement.

At the explicit request of government officials, the company is limiting initial access to a small, pre-approved group of trusted partners whose identities have been shared with federal authorities.

Testing during this restricted preview phase will focus heavily on cybersecurity and safety configurations tailored to each model’s specific threshold.

The company stated it intends to progress toward a broader commercial release in the coming weeks once this initial phase of coordinated testing concludes.

On the technical front, the flagship Sol model introduces new processing features labeled “max” and “ultra” reasoning efforts.

The max setting allocates extended internal time for the model to process complex queries deeply before delivering a response.

The ultra mode introduces an architecture that deploys specialized subagents tasked with accelerating multi-layered operations beyond what a single agent can manage.

Internal testing cited by the firm shows measurable advancements in specialized fields, specifically coding, biology, and cybersecurity infrastructure.

Developer pricing for the new generation operates on a per-million-token basis, scaled across the three sizes.

Sol is priced at $5 for input tokens and $30 for output tokens, while the mid-tier Terra drops to $2.50 for input and $15 for output.

The lightweight Luna model is set at $1 for input and $6 for output tokens.

The release also alters how prompt caching is handled by introducing explicit cache breakpoints and establishing a 30-minute minimum lifespan for cached data.

Under the updated terms for GPT-5.6 and subsequent models, operations that write to the cache are billed at 1.25 times the standard uncached input rate for that model tier.

Operations that read from existing cache files will continue to receive the standard 90 percent discount.

Partners currently holding access are expected to continue operational testing under government visibility as the engineering teams prepare the infrastructure for a wider public launch.